Review three patterns separately, because they use different Workday identities. A pass-through screen and an employee AI chat connector should run as the worker who is signed in. An unattended integration should run as its own integration system user, with only the domains that feed needs. A vendor sentence that says “zero retention” does not change either rule. This packet is the Workday evidence to attach, plus the product claims to demand in writing. It was checked against Workday’s public pages on November 16, 2026.
Kiweely’s coming-next products sit on these patterns. Kiweely Leaf is the pass-through screen. Kiweely Bridge is the employee chat connector. Kiweely Roots is the integration platform that says it keeps nothing after a run. All three are coming next. The controls below are Workday’s, and they apply to any vendor you put in front of the same tenant.
What the reviewer is deciding
| Question | Pass-through screen | Employee AI chat | Unattended integration |
|---|---|---|---|
| QuestionWho is the Workday user on the call? | Pass-through screenThe signed-in employee | Employee AI chatThe signed-in employee | Unattended integrationA dedicated integration system user, not an administrator |
| QuestionTypical grant | Pass-through screenAuthorization Code Grant, scoped functional areas | Employee AI chatAuthorization Code Grant, scoped functional areas | Unattended integrationThe integration’s own client and security group |
| QuestionWhat “too much access” looks like | Pass-through screenOne service account reused for every employee | Employee AI chatA shared account, or writes the chat can fire alone | Unattended integrationGet and Put on domains the file never uses |
| QuestionWhat Workday still enforces | Pass-through screenThat worker’s domains, rows, and context | Employee AI chatThat worker’s domains, rows, and context | Unattended integrationOnly the policies you granted that user, after activation |
| QuestionWhat the vendor must prove | Pass-through screenNo HR records kept after the request | Employee AI chatNo HR records or grants kept at rest | Unattended integrationNo records kept after the run; a client scoped to that feed |
Workday documents the employee pattern as user-as-self. On the Live Data Query setup page, users are prompted to log in to Workday, and that method “enforces Workday’s full security model, including row level controls and contextual security, based on the permissions of the user who authorizes the connection.” The same page registers the client with Authorization Code Grant, Bearer access tokens, a redirect URI, a refresh-token timeout left at the default, and only the functional areas you want that catalog to expose. Save the client id, the client secret, and both endpoints when the task shows them. Workday tells you that you will not be able to view all of it later.
That page is about Live Data Query, not about chat or custom screens. The security property it states is still the one to demand: the call runs as the person who signed in, so Workday applies that person’s row and contextual security. A screen or a chat connector that uses one integration system user for the whole company does not have this property, even if the vendor redraws permissions in its own app.
Employee sign-in, not a shared account
Register API Client is secured by the Security Administration domain. Workday’s Time Kiosk instructions show a narrow client: Authorization Code Grant, Proof Key for Code Exchange, Bearer tokens, one redirect URI, and one functional area. Name the client for one product and one tenant. Add a functional area only when a test call as a real worker fails without it.
Workday Studio’s OAuth setup adds two constraints for an interactive client. The redirect URI must be reachable and encrypted by HTTPS. People enrolled in Workday multifactor authentication must complete a second factor when they connect from Studio. They cannot fall back to basic authentication for that connection. Omitting a relevant functional area makes that Studio authentication fail. A client that asks for every functional area “so we do not have to come back” is a finding.
Workday stores OAuth tokens and token-validity settings in the Workday Object Management System. The Agent Security FAQ says Workday cannot detect a compromised credential for you. If you suspect compromise, disable that OAuth 2.0 client. Replay of a stolen access token is possible until it expires. For integration clients, access-token and refresh-token expiry is configurable. For external agents registered through Workday’s agent features, access-token expiry customization is not available; Workday says that is planned for a later release. Do not paste the external-agent limit onto every API client, and do not assume a third-party chat connector is one of those agents. Ask which grant type and which user the call uses.
Agent System Users are not integration system users. The same FAQ says each Workday agent has its own agent system user, stored in the Object Management System, with a dedicated OAuth client, so the agent has an identity you can audit. Those accounts are tied to skills you configure in Agent System of Record. An employee chat connector that is only a user-delegated API client is a different design: there is no agent system user in the middle, and Workday evaluates the employee. Put the vendor’s diagram next to this paragraph and mark which box is the Workday user.
Unattended integrations
Workday’s own integration instructions use a different identity. The Cloud Connect for Learning setup says every such integration needs an integration system user and an integration system security group, that those users must be unique across integrations, and that it is Workday policy to use an integration system user with only the appropriate access. Administrators should not be the account the integration runs as. Select Do Not Allow UI Sessions so that user cannot sign in through the Workday interface. Then grant domains, then run Activate Pending Security Policy Changes.
The Securing Connectors course is the method, and it is not specific to one template. Domain security policies secure most web service operations. Business process security policies secure some Put operations, because those operations start a business process. A new integration security group is on no policies until you add it, and the integration fails until you do. Open the event audit, find the operation, and run View Security for Securable Item. Grant Get for a read. Grant Put or Modify only for a write or a launch. The course says Delivery Assurance will raise integrations with more access than they need.
Adding the group does nothing until someone runs Activate Pending Security Policy Changes, with a comment naming who signed off and the Confirm box selected. Membership changes take effect immediately and do not need activation. Editing policies and activating them can be different people: Security Configuration versus Security Activation. After activation, compare output. A successful run can still drop fields. Test connector field overrides, and remember calculated-field values follow the domains of the underlying fields.
A zero-retention claim does not shrink this list. If the platform moves payroll, a carrier file, or an identity update, the Workday user on that run still needs the domains for those fields, and it should not have the rest.
What “nothing kept” has to mean in the packet
Workday does not certify a vendor’s disk. These are the statements to put in the review, each marked as vendor evidence or as a Workday fact.
Workday facts you can cite. User-as-self applies the authorizing user’s security, including row-level and contextual rules, on the Live Data Query page. OAuth material that Workday itself holds sits in the Object Management System. Disabling the OAuth client is the customer control when a grant is suspect. Security-policy edits stay pending until activation. An integration system user should be unique and limited, and should not have a Workday UI session, on the Cloud Connect page’s statement of policy.
Vendor statements you attach, not Workday statements. For a screen, ask what is stored after the response is sent. For chat, ask where the grant lives when the worker is idle, whether prompts and answers are logged, and whether a write can be submitted from the transcript alone. For integrations, ask whether anything remains except times, counts, and outcomes.
Kiweely’s pages make those claims and are coming next, so treat them as design claims to test, not as a finished control. Leaf: calls run as the signed-in employee; HR data is not kept at rest; stored items are layouts, branding, and connection settings. Bridge: user-delegated Authorization Code Grant with PKCE, never an integration system user; read-only until a write is enabled; that write waits for a browser card showing now and after; grants sealed with AES-256-GCM and decrypted only in memory; logs limited to timestamp, tenant id, tool name, HTTP status, and latency. Roots: one API client per integration, scoped to what it reads; data held only during the run; agents build the code and are not on the path of each run. Hosting on those pages is Kiweely’s cloud or, for Bridge and Roots, an environment you run.
If a claim is missing from the vendor’s page, leave it out of the approval. Do not fill the gap with a synonym.
Evidence to attach
- The API client name, grant type, redirect URI, and the functional areas actually selected. One client per product. For integrations, one user per integration.
- A View Security for Securable Item excerpt for each read operation and each write operation the design uses.
- The pending-changes view before activation, and the activation comment that names the approver.
- A test as the employee, as the manager, and as a worker who must be denied. User-as-self is false if those three see the same rows.
- For any write, the business process security policy if the operation starts a process, and the step where a person confirms the change outside the model’s transcript.
- The vendor’s retention note: fields stored, fields logged, and how a suspected stolen grant is disabled. Disabling the Workday client is your control even when the vendor also revokes its own session.
- The tenant name. Do this in a non-production tenant first. Production is the live system of record; a sandbox is typically a copy of production data, so the test still contains real worker data and needs the same handling as production data. Tenant types are covered in the sandbox checklist.
Failure modes
Shared service account behind a pretty screen. The portal signs in once. Every employee then sees whatever that account can see, minus whatever rules the portal rebuilt. Workday’s row security never ran for the employee. Reject the design or require user-as-self.
Policy edited, never activated. The group is on the domain in the editor and absent at runtime. The course is explicit: saved policy changes do not take effect until Activate Pending Security Policy Changes. Membership edits are the exception and take effect immediately, including a bad membership you hoped was still pending.
Get and Put copied from a sample. A read-only feed does not need Put. A chat connector that lists write tools the admin has not enabled will fail a review that Bridge’s own page already sets: disabled tasks should not be offered.
Agent System of Record assumed. A user-delegated client is not an agent system user and does not, by itself, require Workday’s agent agreement. Confirm that with the vendor. Bridge’s page says it does not use Agent System of Record. If the design does use Workday’s agents, switch to the Agent Security FAQ and review agent system users on purpose.
“Zero retention” with a full audit log of the payload. Times, counts, and outcomes are not the same thing as stored WQL rows, names, or pay. Ask for the log schema.
Frequently asked questions
Does user-as-self mean the vendor cannot see the data in transit?
No. Workday applies the employee’s security before it returns fields. The vendor can still see the response while it forwards it. Retention is a separate question. Ask what is written after the response is sent.
Can we rely on Workday to notice a stolen OAuth grant?
No. The Agent Security FAQ says Workday cannot detect compromised credentials on your behalf. Disable the OAuth 2.0 client.
Should chat, screens, and integrations share one API client?
No. Scope and identity differ. A client registered for an employee sign-in, with functional areas chosen for questions, is the wrong client for a payroll file. Workday’s integration guidance wants a unique integration system user per integration, with only the access that integration needs.
Where do Kiweely Leaf, Bridge, and Roots fit?
They are the three coming-next products on these patterns: Leaf for screens, Bridge for employee chat, Roots for integrations that keep nothing after the run. Register interest on each product page. They are not a completed security certification, and this guide does not treat their design claims as already tested in your tenant.
Sources checked
- Set Up Live Data Query User-as-Self Authentication in Workday
- FAQ: Agent Security
- Security Policy Configuration and Activation
- Securing Connectors
- Set Up Integration System User Security for Cloud Connect for Learning
- Register API Clients for Time Kiosks
- Enable Multifactor Authentication in Workday Studio